MolecularRNN: Generating realistic molecular graphs with optimized properties

Abstract

Designing new molecules with a set of predefined properties is a core problem in modern drug discovery and development. There is a growing need for de-novo design methods that would address this problem. We present MolecularRNN, the graph recurrent generative model for molecular structures. Our model generates diverse realistic molecular graphs after likelihood pretraining on a big database of molecules. We perform an analysis of our pretrained models on large-scale generated datasets of 1 million samples. Further, the model is tuned with policy gradient algorithm, provided a critic that estimates the reward for the property of interest. We show a significant distribution shift to the desired range for lipophilicity, drug-likeness, and melting point outperforming state-of-the-art works. With the use of rejection sampling based on valency constraints, our model yields 100% validity. Moreover, we show that invalid molecules provide a rich signal to the model through the use of structure penalty in our reinforcement learning pipeline.Preprint. Under review.

Incorporating domain knowledge into neural-guided search

“…In the field of SO, Popova et al (2019) use valency constraints to follow rules of organic chemistry to optimize molecular structures. Li et al (2019) use a neural-guided approach to identify asymptotic constraints of leading polynomial powers and use those constraints to guide MCTS for the problem of symbolic regression.…”

“…A popular SO problem is neural architecture search (NAS), in which tokens represent architectural hyperparameters, the sequence represents a specification of a neural network architecture, and the reward is the validation accuracy when instantiating the specified network and training it on some downstream task (Zoph and Le, 2016). Other examples include program synthesis (Abolafia et al, 2018), symbolic regression , de novo molecular design (Popova et al, 2019), automated theorem proving (Bibel, 2013), and many traditional combinatorial optimization problems (e.g. traveling salesman problem) (Bello et al, 2016).…”

“…In other cases, τ is not executable but is simply evaluated as a whole by a reward function R(τ ). Examples of fields with non-executable objects include de novo molecular design (Popova et al, 2019), automated theorem proving (Bibel, 2013), and many traditional combinatorial optimization problems (e.g. traveling salesman problem) (Bello et al, 2016).…”

Permutation Equivariant Generative Adversarial Networks for Graphs

“…Some use a canonical sequential representation such as the SMILES representation (for instance: [20,17,5,7,2]). Others are generating directly some graph objects such nodes or edges [25,18,9,21,10,8,6].…”

On Evaluation Metrics for Graph Generative Models

“…Graph generation is a key problem in a wide range of domains such as molecule generation (Samanta et al, 2020;Popova et al, 2019;Li et al, 2018;Kong et al, 2021;Jin et al, 2020) and structure generation (Bapst et al, 2019;Thompson et al, 2020). An evaluation metric that is capable of accurately measuring the distance between a set of generated and reference graphs is critical for advancing research on graph generative models (GGMs).…”